1. Under-Exposed
Gaynor Bennett

2. Do we need to wait so long?
The Investigation
Sample Here
Do we need to wait so long?
2 years ago
3 years ago!

3. Extent & effect of good/bad misclassification
What Will Be Revealed?
Industry case studies
Personal Loans
Motor Finance
Retail
Mortgages
Credit Cards
Analysis samples
Extent & effect of good/bad misclassification
Model & comparison approach
Results
Conclusion

4. Industries – Typical Exposure Periods

5. Industries – Typical Exposure Periods

6. Under-Exposed Samples
Min 6m
Ave 12m 6m

7. Under-Exposed Samples
Min 12m
Ave 18m
12m

8. Under-Exposed Samples
Full Exposure
All samples here
12m 6m

9. Good/Bad Classification
Conventional
Bad = worst status 3+ in arrears
Indeterminate = worst status 2 in arrears
Good = worst status 1 in arrears
Adjusted
Bad = worst status 2+ in arrears
Indeterminate = worst status 1 in arrears
Good = worst status 0 in arrears
Credit Cards Adjusted
Indeterminate = empty

10. Good/Bad Misclassification
For Personal Loans – Full Exposure average 33m
After 6m min exposure (average 12m)
of the ultimate bads
45% 9%
46%
But did the early and ultimate bads look different?
And would this affect predictive strength of criteria?

11. Comparison of Bad Profiles

12. Comparison of Bad Profiles
Early bads worse
but….

13. Comparison of Bad Profiles
Not the case for time-related loan details

14. Impact on Predictive Strength
Criterion strength = Information Value
Based upon proportions of goods and bads across attributes
GB strength

15. Impact on Predictive Strength

16. Impact on Predictive Strength
Predictive Nature = Pattern of Weights of Evidence for attributes
GB attribute strength

17. Predictive Nature - Delphi

18. Predictive Nature – Residential Status & Age

19. Predictive Nature – Loan Amount & Term

20. Comparison of Predictive Strength
Predictive power retained on 3+ definition
Slightly weaker on 2+ definition
some difference in patterns
Looked promising for scorecard modelling
time-related loan details best avoided

21. Model Comparison Approach
Develop scorecards for all samples
FullExp
6m(3+)
6m(2+)
12m(3+)
12m(2+)
Known good/bads only
reject inference similar across samples – but could bias comparison
Same criteria and groupings in scorecards
negligible impact
Minimal manual point adjustment

22. Model Comparison Approach
For each scorecard and sample, compare
Predictive content of criteria
‘Balance’ of scorecards
Discrimination, Gini, KS D = ( ) G B - + m s 2 1
100 x KS
Gini

23. Model Comparison Approach
Assuming that the Full Exposure scorecard and sample were the optimum…
Note performance of the other models based on the Full Exposure sample
Discrimination, Gini, KS
Predicted bad rates
Profile of accepts

24. Personal Loans – Scorecard Strength
33m
Based on own sample, 6m(3+) almost as strong!
Less variation in strength when applied to FullExp
On FullExp sample,
12m cards lost 6% discrimination
6m(3+) lost 8%

25. Personal Loans – Criteria Contribution

26. Personal Loans – Criteria Contribution

27. Personal Loans – Bad Rate Predictions

28. Personal Loans – Accepted Profile

29. Personal Loans – Accepted Profile

30. Personal Loans – Summary
On FullExp sample
12m cards lost 6% discrimination
6m(3+) lost 8%
Bad rate predictions very similar
Accepted profiles acceptable
especially after experienced adjustment!
And for the other portfolios…..

31. Motor – Scorecard Strength
Based on own sample, 6m(3+) strongest!
Similar on FullExp
12m cards lost 2% discrimination
6m(3+) lost 4%

32. Motor – Model Results Comparison
Small variation in ‘balance’ of scorecards
Similar accepted profiles
Same bad rate estimates

33. Retail – Scorecard Strength
24m
Based on own sample, 6m(3+) strongest!
Similar on FullExp
12m(3+) card lost 4% discrimination
6m(3+) lost 5%

34. Retail – Model Results Comparison
Small variation in ‘balance’ of scorecards
Similar accepted profiles
Very similar bad rate estimates

35. Mortgages – Scorecard Strength
Very low sample counts
even FullExp insufficient for optimum model
but comparison still relevant
Based on own sample, all scorecards stronger than FullExp
2+ stronger than 3+
Based on FullExp
6m cards lost 6%
12m(3+) & (2+) cards as effective….

36. Mortgages – Scorecard Strength
Attitude to your mortgage
Of all your loans – this is your prime concern
Try not to miss a payment!
Therefore
2+ is bad
1 is indicative of problems
Adjusted definition is valid

37. Mortgages – Model Results Comparison
Small variation in ‘balance’ of scorecards
Similar accepted profiles
Very similar bad rate estimates

38. Credit Cards – Scorecard Strength
25m
Based on own sample,
6m(3+) as strong as FullExp
2+ noind much weaker
Based on FullExp
3+ cards virtually as effective
2+ cards gained in strength

39. Credit Cards – Model Results Comparison
Small variation in ‘balance’ of scorecards
Similar accepted profiles
Very similar bad rate estimates

40. Conclusions
Under-exposed scorecards only marginally less effective
6m(3+) cards on average 5% weaker, min 1%, max 8%
12m(3+) cards on average 2% weaker, min 0%, max 7%
Predicted bad rate savings virtually the same
Accepted profiles very similar

41. Slight difference in predictive nature On own sample
Conclusions
2+ definition
Slight difference in predictive nature
On own sample
generally weaker
much weaker for credit cards (no indeterminates)
slightly stronger for mortgages!
Applied to Full Exposure sample
nearly as good as 3+ cards

42. Little benefit in waiting over 2 years
Conclusions
Little benefit in waiting over 2 years
But benefit to be gained from Under-Exposure…..
For generic scorecards
earlier access to enhanced bespoke decisions
Even established scorecard users
benefit from a more representative sample
avoid archiving or missing data problems
faster fine-tunes
Monitor at an early stage
or factor up to ultimate bad rates
So why wait? Let’s get….

43. Under-Exposed!
Gaynor Bennett